Watch on YouTube · subscribe for the daily brief

"Yesterday authors told a court OpenAI took their work without asking. Today Universal Music and ElevenLabs announced the opposite: license it, let artists opt in, pay them. Two days, two opposite answers to the same question, and the difference is a single word, consent.

Sometimes the news gives you a perfect natural experiment, and this week did. On September 9, a class of authors told a federal judge that OpenAI built ChatGPT on their work without asking, torrenting books and hiding it. On September 10, Universal Music Group and ElevenLabs announced the exact opposite approach to the exact same problem: license the catalog, let artists choose whether to participate, and pay them when their work is used. Same underlying question, where does AI's training material come from and are the creators compensated, answered two completely different ways in twenty-four hours. That contrast is the story, and it is more useful than either announcement alone.

Here is the deal itself. Universal Music Group, the largest record label in the world, signed a multi-year licensing agreement with ElevenLabs, the AI audio company best known for voice cloning and speech synthesis. Together they will build a new AI music platform on top of UMG's catalog, letting fans create remixes, mashups, new interpretations of tracks, and what the companies call "personalized vocal experiences." Three details make it matter. It is ElevenLabs' first agreement with a major label, so it is a template, not a one-off. Artists opt in, participation is a choice, not a default, and UMG says participating artists get fairly compensated when their work is used. And the framing from both CEOs leans hard on one word: responsible. Grainge called it responsible AI that unlocks new revenue for creators, Staniszewski stressed rights management and fair compensation. Whatever you think of the marketing gloss, the structure is real: permission first, opt-in, paid.

Now hold it against yesterday. The authors' filing accused OpenAI of the opposite on every axis: no permission, no opt-in, no payment, and active concealment. That is the whole spectrum of how AI gets its raw material, laid out across two consecutive days. Take-first, defend-with-fair-use-later, currently being litigated. Or ask-first, license, opt-in, pay, currently being signed. And the timing is not a coincidence, it is a trend crystallizing. This deal landed one day after Warner Music Group announced its own licensing partnership with Suno, the AI music generator Warner had been fighting. UMG itself had already settled its litigation with Udio in 2025 and signed a deal. So the pattern across the music industry is unmistakable: the labels are moving, fast, from suing AI companies to licensing them. The lawsuit era and the licensing era are not sequential. They are happening at the same time, and the music industry is a year or two ahead of publishing in making the turn.

Why does a music-licensing deal belong in a brief for data and ML engineers? Because it is the clearest signal yet of where the whole training-data question is heading, and that determines what you will be allowed to build on. For two years the default assumption in AI development was take-first: scrape it, train on it, argue fair use if anyone complains. That assumption is now under direct legal attack (yesterday) and being actively replaced by a licensed alternative (today). If the licensed model wins, and the music industry is betting real money that it will, then the future of training data is opt-in catalogs with compensation baked in, not scraped piles defended in court. For anyone building AI products, that is the strategic read: the era of "just scrape it" is closing, and the companies structuring their data sourcing around permission and licensing now are the ones who will not be refactoring under a court order later. ElevenLabs did the thing OpenAI is being sued for not doing. That is not just ethics, it is durability.

The honest counterweight, because these deals deserve skepticism too. "Responsible AI" and "fairly compensated" are the labels' words, and the labels have every incentive to frame a deal that monetizes their catalog as a win for artists. The actual terms, how much artists get, how the opt-in really works, whether a working musician sees meaningful money or a rounding error, are not public, and "fair compensation" has meant very little in past music-industry deals. There is also a real question of whether fans actually want to generate remixes and "personalized vocal experiences," or whether this is labels finding a new way to sell the same catalog. So do not read this as artists winning. Read it as the industry choosing the licensed model over the litigated one, which is the structurally important part regardless of whether the artist economics turn out to be generous or thin. The direction is the signal. The terms will decide who actually benefits.

The rest of the day, and there is a genuinely important security item plus a data-engineering standout. Anthropic alleged that Chinese AI firms routed user queries through Claude, per RIA, flag that one as single-source and state-media, so verify before repeating, but if it holds it is another data-provenance story, who is quietly using whose model. On the security front, The Hacker News reported an executive-extortion campaign where attackers impersonate IT help desks to steal Microsoft 365 data, worth knowing because help-desk impersonation is a rising social-engineering vector and your execs are the target. On the builder's bench, the standout for your audience is Estuary's "PostgreSQL CDC Backfills at Scale," running a multi-day backfill while production keeps writing, which is a real, hard, unglamorous data-engineering problem explained well, exactly the kind of thing your readers actually hit. Starburst's "What is Data Lineage?" pairs with it nicely, and lineage is quietly the same theme as the day's big story, knowing where data came from. ClickHouse published another performance-per-dollar comparison against Snowflake, vLLM optimized MiniMax M3 on AMD (the hardware-independence thread continues), and Hugging Face rebuilt AUTOMATIC1111 with Gradio Workflow. Simon Willison flagged "Native is now the future of mobile at Shopify" and a browser-based Nix package tool.

The through-line, one sentence: on September 10 Universal Music and ElevenLabs announced a licensed, opt-in, paid AI music platform, and the reason it matters lands only when you read it against September 9, where authors accused OpenAI of taking their work without any of those things, because together the two days lay out the entire spectrum of AI's training-data future, take-first-and-litigate versus ask-first-and-license, and the money is starting to move toward permission.

One thing to do today: notice which side of the take-first-versus-ask-first line your own AI work sits on, because the industry just showed you both futures in two days and started picking one. Concretely: for any data you train on, fine-tune with, or feed into a model, can you answer "do we have the right to use this," cleanly, the way ElevenLabs can for UMG's catalog, or is your answer closer to "we scraped it and we'd argue fair use," the way OpenAI is arguing in front of Judge Stein right now? The licensed model is winning momentum, the music industry is betting on it a full year ahead of everyone else, and the teams sourcing their data around permission today are the ones who will not be rebuilding under legal pressure tomorrow. Permission is turning into a business model. Get on the right side of that line before a court, or a licensing market, moves it for you.

Sources covered

1 story · 13 sources (3 news, 7 company, 3 voices)

1. Any Nix package, live in your browser
   🎙️Simon Willison↗🎙️Simon Willison↗📰Engadget↗📰RIA Novosti↗🏢Estuary↗🏢Starburst↗🏢Snowflake↗🏢Diraq↗📰The Hackers News↗🎙️Simon Willison↗🏢vLLM↗🏢Hugging Face↗🏢ClickHouse↗